Closer to actually applying matrices.

inducer · Jul 30, 2009 · 6afce29 · 6afce29
1 parent a627a1e
commit 6afce29
Show file tree

Hide file tree

Showing 7 changed files with 38 additions and 65 deletions.
diff --git a/CMake/FindMETIS.cmake b/CMake/FindMETIS.cmake
@@ -30,8 +30,7 @@ FIND_LIBRARY(METIS_LIBRARY metis
   /usr/lib
   )
 
-MESSAGE("${METIS_INCLUDE_DIR} yo")
-MESSAGE("${METIS_LIBRARY} yo")
+MARK_AS_ADVANCED(METIS_INCLUDE_DIR METIS_LIBRARY)
 
 IF(METIS_INCLUDE_DIR AND METIS_LIBRARY)
   SET( METIS_FOUND "YES" )
@@ -40,5 +39,3 @@ ELSE()
     MESSAGE(FATAL_ERROR "METIS not found")
   ENDIF()
 ENDIF()
-
-MARK_AS_ADVANCED(METIS_INCLUDE_DIR METIS_LIBRARY)
diff --git a/CMakeLists.txt b/CMakeLists.txt
@@ -7,6 +7,8 @@ set(CMAKE_MODULE_PATH
 
 project(iterative_cuda)
 
+set(CUDA_ARCH "compute_13" CACHE STRING "Target CUDA architecture")
+
 find_package(CUDA QUIET REQUIRED)
 find_package(METIS REQUIRED)
 

diff --git a/include/iterative-cuda.hpp b/include/iterative-cuda.hpp
@@ -64,6 +64,7 @@ namespace iterative_cuda
       void to_cpu(value_type *cpu);
 
       value_type *ptr();
+      const value_type *ptr() const;
   };
 
 
@@ -100,16 +101,10 @@ namespace iterative_cuda
       index_type row_count() const;
       index_type column_count() const;
 
-      void permute(
-          vector_type const &dest,
-          vector_type const &src) const;
-      void unpermute(
-          vector_type const &dest,
-          vector_type const &src) const;
+      void permute(vector_type &dest, vector_type const &src) const;
+      void unpermute(vector_type &dest, vector_type const &src) const;
 
-      void operator()(
-          vector_type const &dest,
-          vector_type const &src) const;
+      void operator()(vector_type &dest, vector_type const &src) const;
   };
 
 

diff --git a/src/CMakeLists.txt b/src/CMakeLists.txt
@@ -1,11 +1,10 @@
 set(BUILD_SHARED_LIBS ON)
 
 include_directories(${METIS_INCLUDE_DIR})
-message("${METIS_LIBRARY} bitches")
 
 cuda_add_library(iterativecuda
-  host_instantiation.cpp 
   instantiation.cu
+  OPTIONS "-arch=${CUDA_ARCH}"
   )
 
 target_link_libraries(iterativecuda ${METIS_LIBRARY})

diff --git a/src/gpu-sparse-matrix.hpp b/src/gpu-sparse-matrix.hpp
@@ -35,6 +35,7 @@ SOFTWARE.
 #include "helpers.hpp"
 #include "spmv/partition.h"
 #include "spmv/csr_to_pkt.h"
+#include "spmv/utils.h"
 
 
 
@@ -69,7 +70,7 @@ namespace iterative_cuda
     csr_mat.Aj = const_cast<index_type *>(csr_column_indices);
     csr_mat.Ax = const_cast<value_type *>(csr_nonzeros);
 
-    index_type rows_per_packet = 
+    index_type rows_per_packet =
       (SHARED_MEM_BYTES - 100)
       / (2*sizeof(value_type));
 
@@ -78,7 +79,10 @@ namespace iterative_cuda
     std::vector<index_type> partition;
     partition_csr(csr_mat, block_count, partition, /*Kway*/ true);
 
-    pimpl->matrix = csr_to_pkt(csr_mat, partition.data());
+    pkt_matrix<index_type, value_type> host_matrix =
+      csr_to_pkt(csr_mat, partition.data());
+
+    pimpl->matrix = copy_matrix_to_device(host_matrix);
   }
 
 
@@ -87,44 +91,49 @@ namespace iterative_cuda
   template <typename VT, typename IT>
   gpu_sparse_pkt_matrix<VT, IT>::~gpu_sparse_pkt_matrix()
   {
-    delete_pkt_matrix(pimpl->matrix);
+    delete_pkt_matrix(pimpl->matrix, DEVICE_MEMORY);
   }
 
 
 
 
   template <typename VT, typename IT>
-  gpu_sparse_pkt_matrix<VT, IT>::index_type
-  gpu_sparse_pkt_matrix<VT, IT>::row_count() const
-  { 
+  IT gpu_sparse_pkt_matrix<VT, IT>::row_count() const
+  {
     return pimpl->matrix.num_rows;
   }
 
 
 
 
   template <typename VT, typename IT>
-  gpu_sparse_pkt_matrix<VT, IT>::index_type
-  gpu_sparse_pkt_matrix<VT, IT>::column_count() const
-  { 
+  IT gpu_sparse_pkt_matrix<VT, IT>::column_count() const
+  {
     return pimpl->matrix.num_cols;
   }
 
+
+
+
   template <typename VT, typename IT>
   void gpu_sparse_pkt_matrix<VT, IT>::permute(
-      vector_type const &dest,
+      vector_type &dest,
       vector_type const &src) const
   {
+    gather_device(dest.ptr(), src.ptr(), 
+        pimpl->matrix.permute_old_to_new, row_count());
   }
 
 
 
 
   template <typename VT, typename IT>
   void gpu_sparse_pkt_matrix<VT, IT>::unpermute(
-      vector_type const &dest,
+      vector_type &dest,
       vector_type const &src) const
   {
+    gather_device(dest.ptr(), src.ptr(), 
+        pimpl->matrix.permute_new_to_old, row_count());
   }
 }
 

diff --git a/src/host_instantiation.cpp b/src/host_instantiation.cpp
diff --git a/src/instantiation.cu b/src/instantiation.cu
@@ -26,8 +26,18 @@ SOFTWARE.
 
 
 #include <iterative-cuda.hpp>
+#include "gpu-vector.hpp"
+#include "gpu-sparse-matrix.hpp"
 
 
 
 
 using namespace iterative_cuda;
+
+
+
+
+template class gpu_vector<float>;
+template class gpu_vector<double>;
+template class gpu_sparse_pkt_matrix<float>;
+template class gpu_sparse_pkt_matrix<double>;